On Thu, Sep 07, 2023 at 11:23:40AM +0200, Uladzislau Rezki wrote:
On Thu, Sep 07, 2023 at 08:11:48AM +0100, Lorenzo Stoakes wrote:
[..]
Anyway, so TL;DR:-
- As we both agree, add a comment to explain why you need the spin trylock.
(there are no further steps :P)
And I don't believe this actually needs any further changes after this discussion*, so if you fancy doing a follow up to that effect that will suffice for me thanks!
Thanks.
For PREEMPT_RT kernels we are not allowed to use "vmap parts" in non slepable context, this is just reality, because we use a sleep type of spinlock.
I am not sure how urgent we need this fix. But to me it looks like debuging and corner case. Probably i am wrong and miss something. But if it is correct, i would just bailout for RT kernel and rework later in a more proper way. For example implement a safe way of RCU scan but this is also another story.
Bailing out for RT kernel is insufficient, as we need the trylock() to avoid self-deadlock as well for !PREEMPT_RT. Plus IIRC in the past there was a opposition to special-casing PREEMPT_RT in code as well. Admittedly those PREEMPT_RT cases were related to detecting preempt-disabled than a lock-held section though.
We could optionally do a trylock() loop + bail out after certain number of tries as well but that would compilicate the code a bit more and I am not sure if it is worth it. Still if you guys feel strongly about doing something like that, let me know and I can give it a try :).
thanks,
- Joel